人工智能模型检测牙根纵裂有效性的系统综述
该系统综述评估了人工智能模型在不同影像技术下检测牙根纵裂的诊断性能,检索截至2025年1月。纳入7项研究,卷积神经网络模型准确率75%-97.8%、敏感度75%-98%、特异度60%-99%,ResNet-50应用于手动裁剪CBCT切片准确率达97.8%。证据确定性低,需前瞻性多中心研究验证。
要点
- CNN模型检测牙根纵裂准确率75%-97.8%,基于CBCT的模型优于2D影像。
- ResNet-50用于手动裁剪CBCT切片时准确率最高达97.8%。
- 证据确定性低,受方法学局限、样本量小和外验证不足限制。
核心结论
- 低水平证据表明基于卷积神经网络的AI模型,尤其经高分辨率CBCT和增强图像训练后,检测牙根纵裂可达高诊断准确率。
研究设计与样本
- 系统综述,检索PubMed、Scopus和Web of Science至2025年1月,初检1544项研究,最终纳入7项。
- 纳入报告诊断准确率、敏感度和特异度的研究,排除非英语文献,采用QUADAS-2评价质量,GRADE评级证据确定性。
主要发现
- 卷积神经网络模型准确率75%-97.8%,敏感度75%-98%,特异度60%-99%。
- ResNet-50应用于手动裁剪CBCT切片准确率最高达97.8%。
- 概率神经网络和集成架构在大型平衡数据集训练时表现良好。
- 使用全景或根尖片等低分辨率模态或自动感兴趣区选择时性能下降。
- 基于CBCT训练的模型一致优于基于2D放射摄影的模型。
临床怎么用
- AI模型可作为牙根纵裂的潜在诊断辅助工具,优先考虑基于高分辨率CBCT训练的模型。
- 避免在低分辨率影像或自动感兴趣区选择条件下依赖模型性能。
- 临床实施前需经前瞻性多中心研究验证泛化能力。
局限性
- 证据确定性低,存在方法学局限、样本量小和外部验证有限的问题。
- 需前瞻性、多中心研究使用临床采集数据集确认泛化性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
INTRODUCTION: Vertical root fractures (VRFs) are diagnostically challenging lesions with significant clinical implications. Artificial intelligence (AI) models have emerged as a potential diagnostic aid. This systematic review assessed the diagnostic performance of AI-based models for VRF detection across various imaging techniques. METHODS: Databases including PubMed, Scopus, and Web of Science were thoroughly searched until Jan 2025. Studies that reported on diagnostic accuracy, sensitivity, and specificity were included. Articles in languages other than English were excluded. Study quality was evaluated using QUADAS-2, and the certainty of evidence was rated following the GRADE approach. RESULTS: Of the initial 1,544 studies, 7 met the inclusion criteria. Across these studies, convolutional neural network-based models showed a 75%-97.8% accuracy, 75%-98% sensitivity, and 60%-99% specificity. Applying ResNet-50 to manually curated cone-beam computed tomography (CBCT) slices achieved the highest accuracy at 97.8%. Probabilistic neural networks and ensemble-based architectures also performed well, especially when trained on large, balanced datasets. Performance declined when lower-resolution modalities such as panoramic or periapical radiographs were used, or when automatic region-of-interest selection was applied. Models trained on CBCT consistently outperformed those using 2D radiography. CONCLUSIONS: There is low-level evidence that indicates convolutional neural network-based AI models, especially when trained on high-resolution CBCT and enhanced images, can achieve high diagnostic accuracy for VRF detection. The overall certainty of this evidence remains low due to methodological limitations, small sample sizes, and limited external validation. Prospective, multicenter studies using clinically acquired datasets are necessary to confirm generalizability and support clinical implementation.
引用
- Patil S, Licari FW, Bennion C, et al. Effectiveness of Artificial Intelligence Models for Detection of Vertical Root Fractures: A Systematic Review. J Endod. 2026. PMID: 41740650. doi:10.1016/j.joen.2026.02.012
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。